Port ERNIE-4.5-21B-A3B-Base (ernie4_5_moe)

[DarthCeltic]

Summary

• Adds baidu/ERNIE-4.5-21B-A3B-Base-PT to the most_models_ported track (identity ernie4_5_moe), introducing the ERNIE 4.5 MoE execution family to the board (21B total params / 3B active per token — "A3B", 64 routed experts + 2 shared experts, 28 layers, 131K context).
• Distinct from the already-claimed ernie45_03b identity (PR feat(llama_cpp_et): add ERNIE-4.5-0.3B, CodeShell-7B-Chat, Cohere2 causal-LM ports #185): that port covers ERNIE-4.5-0.3B, a small dense model via the separate Ernie4_5Model converter class. This port targets Ernie4_5MoeModel (a subclass), a genuinely different code path — expert routing, shared experts, different tensor shapes — and a distinct architecture/identity.
• No converter fix was needed: Ernie4_5MoeModel is already correctly registered in this repo's convert_hf_to_gguf.py, so this converted cleanly on the first attempt with the stock, unmodified converter — same clean-conversion category as smallthinker and afmoe.
• Self-converted to Q8_0 GGUF (389 tensors, 23.2 GB, sha256=2d18e66097f598c8fb5f3d15fd2fba9e6f916652985426bf0e7e1b478a625d48), hosted on Hugging Face (darthceltic85/ernie45-21b-a3b-base-gguf) since it far exceeds GitHub's 2 GB release-asset limit.
• Verified live: model loads cleanly on the CPU backend (arch = ernie4_5-moe, 64 experts / 6 used + shared experts, correct MoE tensor layout, clean compute graph) and real llama-perplexity against WikiText-2 raw gives PPL = 6.8926 +/- 1.25185 — a good, in-range result for a 21.83B-parameter model.

Test plan

• benchmark_config in the claim JSON exactly matches .github/ci/benchmark_config.json's new ernie45_21b_a3b_base entry
• Source revision is a valid 40-char hex commit SHA
• RECIPE.md exists and documents both the clean conversion and the distinction from the existing dense ERNIE-4.5 identity
• All new/modified JSON files are valid JSON
• Model loads and produces a real, in-range perplexity result (not fabricated)